Transcription of arXiv:1702.08502v3 [cs.CV] 1 Jun 2018
{{id}} {{{paragraph}}}
Example: biology
arXiv:1702.08502v3 [cs.CV] 1 Jun 2018
understanding Convolution for Semantic SegmentationPanqu Wang1, Pengfei Chen1, Ye Yuan2, Ding Liu3, Zehua Huang1, Xiaodi Hou1, Garrison Cottrell41 TuSimple,2 Carnegie Mellon University,3 University of Illinois Urbana-Champaign,4UC San advances in deep learning, especially deep con-volutional neural networks (CNNs), have led to significantimprovement over previous semantic segmentation we show how to improve pixel-wise semantic seg-mentation by manipulating convolution-related operationsthat are of both theoretical and practical value. First, wedesign dense upsampling convolution (DUC) to generatepixel-level prediction, which is able to capture and decodemore detailed information that is generally missing in bi-linear upsampling. Second, we propose a hybrid dilatedconvolution (HDC) framework in the encoding phase.
in image understanding and self-driving systems. The re-cent success of deep convolutional neural network (CNN) models [17, 26, 13] has enabled remarkable progress in pixel-wise semantic segmentation tasks due to rich hier-archical features and an end-to-end trainable framework [21, 31, 29, 20, 18, 3]. Most state-of-the-art semantic seg-
Tags:
Information
Domain:
Source:
Link to this page:
Please notify us if you found a problem with this document: